How design rounds are scored, by level
The same prompt, three bars: Senior builds it, Staff chooses between ways to build it, Principal decides whether and how the organisation should.
Neither Anthropic nor OpenAI publishes a design rubric. OpenAI's guide says it looks for "well-designed solutions" and strong collaboration (official, Sources O3); third-party reports of both labs describe a dedicated system design round (Sources A9, O9). The rubric below is ours, built from those signals and from common industry practice. Treat it as a way to practise, not as either company's scoresheet.
| Dimension | Senior | Staff | Principal |
|---|---|---|---|
| Problem framing | Asks good clarifying questions | Narrows scope deliberately and says what is out | Reframes: is this the right problem, for which users, at what cost of delay? |
| Requirements | Lists functional and non-functional | Quantifies them (QPS, p99, durability, budget) and ranks them | Ties them to business outcomes and to what the org can operate |
| Design | A coherent design that works | Two or more options with a reasoned choice | A target architecture plus the path to it from today's system |
| Trade-offs | Mentions some | Names what each choice costs and when it would flip | Weighs reversibility, team topology and long-term cost of ownership |
| Depth | Goes deep where asked | Picks the risky component and goes deep unprompted | Goes deep on the decisions that are expensive to reverse |
| Operations | Mentions monitoring | SLOs, failure modes, rollout, on-call | Org-level: ownership boundaries, migration risk, governance |
| Communication | Clear | Drives the session; checks in at decision points | Makes the interviewer a collaborator; summarises decisions and open risks |
The 45-minute shape
- Requirements and scope (5 to 8 min). Users, core flows, scale, latency and durability targets. Write them where the interviewer can see them.
- Estimates (3 to 5 min). Only the numbers that change the design: peak QPS, storage growth, fan-out. See the next lesson.
- API and data model (5 min). The two or three endpoints and the main entities. This is where most designs quietly go wrong.
- High-level design (10 min). Boxes and arrows for the main flows. Name the storage choice and why.
- Deep dives (15 min). The one or two components where the risk is. Offer a choice: "I'd like to go deep on the fan-out or on consistency; which is more useful to you?"
- Wrap-up (3 min). Bottlenecks, what you would build next, what you would measure first.
Practice
Written for this site to drill the lesson. Not reported interview questions.
- Pick any worked prompt in this track. Write the requirements section only, in five minutes, with numbers.
- Take a system you built. Write the Staff-column version of its trade-offs: three decisions, the rejected alternative for each, and what would flip it.
- Rehearse the wrap-up: in 90 seconds, summarise a design, its two biggest risks and what you would measure first.
Check yourself
What separates the Staff column from the Senior column on design?
Two or more options with a reasoned choice, and naming what each choice costs and when it would flip.
What does a Principal answer add?
The path from today's system to the target, reversibility, team ownership and long-term cost of ownership.
Sources
- Official OpenAI interview guide